This study examines the forecasting performance of time varying coefficient regression models under alternative structures of coefficient variation. It compares the forecasting performance of fixed and stochastic coefficient models when the true Data Generating Process includes either constant or varying coefficients with stationary and I (1) regressors. Furthermore, it investigates the cost of misspecification in the coefficients structure and it tries to find which structure provides the smallest deterioration in forecast accuracy. Besides, it examines the ability of recursive coefficient plots, Cusum and Cusum Square test to correctly advice for parameter instability. Finally, it uses the stochastic coefficient regression models in order to improve the out-of-sample forecasts of two structural macroeconomic models, the Frenkel-Bilson and Dornbusch-Frankel, for the case of the Dollar/Euro Exchange Rate.